Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution

Monishwaran Maheswaran; Leon Lakhani; Zhongzhu Zhou; Shijia Yang; Junxiong Wang; Coleman Hooper; Yuezhou Hu; Rishabh Tiwari

cs.AIApr 9, 2026

Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution

Authors: Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou, Shijia Yang, Junxiong Wang, Coleman Hooper, Yuezhou Hu, Rishabh Tiwari,

Organizations: UC Berkeley · Together AI · UT Austin · Princeton University ·

Abstract

We show that verifier-free evolution is bottlenecked by both diversity and efficiency: without external correction, repeated evolution accelerates collapse toward narrow modes, while the uniform use of a high-cost model wastes compute and quickly becomes economically impractical. We introduce Squeeze Evolve, a unified multi-model orchestration framework for verifier-free evolutionary inference. Our approach is guided by a simple principle: allocate model capability where it has the highest marginal utility. Stronger models are reserved for high-impact stages, while cheaper models handle the other stages at much lower costs. This principle addresses diversity and cost-efficiency jointly while remaining lightweight. Squeeze Evolve naturally supports open-source, closed-source, and mixed-model deployments. Across AIME 2025, HMMT 2025, LiveCodeBench V6, GPQA-Diamond, ARC-AGI-V2, and multimodal vision benchmarks, such as MMMU-Pro and BabyVision, Squeeze Evolve consistently improves the cost-capability frontier over single-model evolution and achieves new state-of-the-art results on several tasks. Empirically, Squeeze Evolve reduces API cost by up to $\sim$ 3 $\times$ and increases fixed-budget serving throughput by up to $\sim$ 10 $\times$ . Moreover, on discovery tasks, Squeeze Evolve is the first verifier-free evolutionary method to match, and in some cases exceed, the performance of verifier-based evolutionary methods.

Figures & tables

Figure 1: Squeeze Evolve shifts the cost–capability frontier left by combining verifier-free evolution with multi-model orchestration. Left: Conceptual scaling curves. Right: Key results across ARC-AGI-V2, MMMU-Pro, and BabyVision.

Method	$k$	$\|\mathcal{P}\|$	$\mathrm{select}$	$\mathrm{recombination}$	Fitness $f$	Model
Majority Voting	1	$N$	Answer clustering	Identity	Consensus frequency	Single
Self-Refinement	$T$	1	Self-critique	Conditioned rewrite	Natural language critique	Single
RSA	$T$	$N$	$K$ -subset	LLM aggregation	Implicit	Single
AlphaEvolve	$T$	Variable	Fitness-guided	LLM aggregation	External $h$	Multi-model
Squeeze Evolve	$T$	Variable	Fitness-guided	Mix of recombination	Probabilistic fitness	Multi-model

Table 1: Instantiation of the unified evolutionary framework for test-time scaling methods.

Figure 2: Single-model open-loop evolution collapses diversity and shrinks the population’s pass@

K

ceiling, while multi-model routing preserves both. GPQA-Diamond results across RSA loops in two comparison settings. In both settings, the single-model baselines lose diversity after the early loops and show a corresponding decline in pass@

K

, whereas Squeeze Evolve remains higher and flatter on both metrics. Shaded bands show variation across seeds.

Model pair	Data	S $\rightarrow$ W	W $\rightarrow$ S	$\Delta$
GPT-OSS-120B / GPT-OSS-20B	HMMT’25	0.89	0.85	+4
Qwen3-4B-Thinking-2507 / Qwen3-4B-Instruct-2507	AIME’25	0.88	0.65	+23

Table 2: Ancestor function dominates final accuracy. Mean final loop (9) accuracy across 4 seeds. Strong-init

\rightarrow

weak-agg outperforms weak-init

\rightarrow

strong-agg, indicating initialization quality dominates aggregation quality.

Panel of Figure 3 — Figure 3: Aggregation success is seed-dependent, and group confidence predicts it. (a) With zero correct seeds, neither model recovers a correct answer; with all seeds correct, both achieve near-perfect accuracy. Full results across all seed counts (0–4) in Appendix Figure . (b) Mean group confidence (GC) across RSA loops 1–9 on AIME 2025, split by whether the subset contains $\geq$ 1 correct trajectory. In both self-model and cross-model settings, correct-containing subsets maintain consistently higher GC ( $\overline{\Delta}\geq+3.0$ ). Full results in Appendix Figures and .

Figure 4: Squeeze Evolve overview. The expensive Model 2 generates the initial population; subsequent loops recombine groups using Model 1 and 2 based on group confidence

Setting	Fitness $f$	$\mathrm{Select}_{f}$	$\mathrm{Recomb}_{f}$ (routing rule)	$\mathrm{Update}$
Math / Coding / Vision	GC (Eq. 4 )	Uniform	Percentile on GC	Replace
ARC-AGI-V2 (§ 6.3 )	Diversity $D$ (Eq. 5 )	Uniform	Threshold on $D$ + lite agg	Replace
Circle Packing (§ 6.4 )	GC (Eq. 4 )	Fitness-weighted ( $\zeta{=}0.5$ )	Percentile on GC	Accumulate

Table 3: Operator instantiations of Squeeze Evolve across evaluation settings.

Panel of Figure 5 — Figure 5: Accuracy vs. cumulative cost for homogeneous model pairs. Each point corresponds to one RSA loop (0–9). Squeeze Evolve (green) tracks the RSA accuracy curve while staying significantly further left, achieving comparable quality at 1.4–2.0 $\times$ lower cost.

Panel of Figure 6 — Figure 6: Accuracy vs. cumulative cost for heterogeneous model pairs. Squeeze Evolve (green) matches the expensive curve at 1.4–1.9 $\times$ lower cost, demonstrating that confidence-based routing generalizes across model families and access types.

Figure 7: Accuracy vs. cumulative cost on MMMU-Pro for homogeneous and heterogeneous vision pairs. Left: Kimi-2.5 Instant (Model 1) / Thinking (Model 2). Right: Qwen3.5-35B (Model 1, text-only)

\to

Kimi-2.5 Thinking (Model 2). Savings are measured at matched accuracy. The heterogeneous pair achieves 2.7

\times

savings despite Model 1 never seeing any images.

Strategy	Model 1	Model 2	Acc.	$/Task	Savings
Code-execution methods
Imbue †	—	Gemini 3.1 Pro	95.1	$8.71	—
Confluence Lab †	—	—	97.9	$11.77	—
Full pipeline ( $T{=}10$ )
RSA	—	Gemini 3.1 Pro	93.3	$28.85	1.0 $\times$
Squeeze Evolve	—	Gemini 3.1 Pro	97.5	$7.74	3.7 $\times$

Table 4: ARC-AGI-V2 public evaluation results.

N{=}4

K{=}2

. † Uses code execution and program synthesis. Extended results in Table (Appendix).

Method	Model	Score ( $\uparrow$ )
ShinkaEvolve ( Lange et al., 2025 )	Ensemble (see caption)	2.635982
Squeeze Evolve	GPT-OSS-120B + 20B	2.635896
AlphaEvolve ( Novikov et al., 2025 )	Gemini-2.0 Pro + Flash	2.635862
OpenEvolve ( Sharma, 2025 )	Gemini-2.0 Flash + Claude 3.7 Sonnet	2.634292

Table 5: Comparison of methods on Circle Packing (

n{=}26

\uparrow

). ShinkaEvolve uses an ensemble of Claude Sonnet-4, GPT-4.1, GPT-4.1-mini, GPT-4.1-nano, and o4-mini.

Figure 8: Routing overhead is minimal. Routing adds 1.9–6.8% to end-to-end latency for the Qwen models and 2.8–12.4% for GPT-OSS-120B, with the higher relative overhead on GPQA reflecting its short absolute generation time (106s). Full results in Table (Appendix).

Figure 9: Fixed-budget throughput speedup over RSA. Under the same total GPU budget, the Qwen pair achieves 4–10

\times

speedup and the GPT-OSS pair 1.4–3.4

\times

. Full results in Table (Appendix).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Benchmark	Size	Answer Format	Metric
AIME 2025	30	Integer (000–999)	Accuracy
HMMT Feb. 2025	30	Short answer	Accuracy
GPQA-Diamond	198	4-way MC	Accuracy
LiveCodeBench V6	175	Code	Pass@1
MMMU-Pro	1,730	Up to 10-way MC	Accuracy
BabyVision	388	Short answer	Accuracy

Appendix

Table 6: Summary of evaluation benchmarks.

Model	Effort	Temp.	Top-K	Top-P	Min-P	Gen. Len.
Qwen3-4B-Instruct-2507	—	0.7	20	0.8	0	8K
Qwen3-30B-A3B-Instruct-2507	—	0.7	20	0.8	0	8K
Qwen3-235B-A22B-Instruct-2507	—	0.7	20	0.8	0	16K
Qwen3-235B-A22B-Thinking-2507	—	0.6	20	0.95	0	32K
Qwen3-30B-A3B-Thinking-2507	—	0.6	20	0.95	0	32K
Qwen3-4B-Thinking-2507	—	0.6	20	0.95	0	32K

Appendix

Table 7: Generation hyperparameters for each model.

Inference-Time Optimization

Inference-time optimization adapts model computation, outputs, or runtime settings to improve quality under deployment constraints.

See all 96 papers

Verifier-Harness Co-Evolution for Training-Free Self-Improvement on Open-Ended Tasks
Oct 13, 2026Tianyi Xu, Yaolun Zhang, Yiran Wu +6LLM Agent Self-Improvement LLM Self-Refinement
Oregon State University · AG2 AI · Pennsylvania State University +1
SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery
Oct 20, 2026Ha Minh Hieu, Tongyao Zhu, Do Xuan Long +1Large Language Model-Guided Scientific Discovery LLMs for Evolutionary Optimization
Hanoi University of Science and Technology · National University of Singapore
FreeEvolve: Learning to Evolve Beyond Fixed Loops
Oct 6, 2026Lecheng Kong, Like Hui, Nikos Kanakaris +3Self-Evolving Agents LLM Agent Self-Improvement
AWS AI Labs · Georgia Institute of Technology

Oct 13, 2026J/K move · Enter open · S save

Verifier-Harness Co-Evolution for Training-Free Self-Improvement on Open-Ended Tasks

Tianyi Xu, Yaolun Zhang, Yiran Wu, Joe Nguyen+5

Oregon State University · AG2 AI · Pennsylvania State University+1

LLM Agent Self-Improvement LLM Self-Refinement Agent Harness Optimization

Open paper AskPDF arXiv

Training-free self-improvement lets an agent use feedback to revise the response it is about to return and to update the harness state that shapes how it handles later tasks. On open-ended tasks that feedback rarely comes from reference answers or executable tests, so the agent has to rely on a model-based verifier. When the verifier is wrong, the current response can get worse, and misleading experience can persist across the task stream. A verifier that never changes also grows stale as the agent's responses improve and task demands shift. We introduce Unified Co-Evolution (UCoE), a training-free method that evolves an executable verifier program together with the agent's harness state while all model parameters stay fixed. UCoE synthesizes candidate verifier programs and tests them on response contrasts, pairs of responses that differ in one declared property and have a known preferred side. It selects the candidate whose discrimination on its weakest contrast family is highest, a quantity we call driving capacity. The selected verifier diagnoses one weakness in each draft, guides a single repair, decides which response is served, and determines whether the accepted repair may enter the harness state. As tasks complete, their diagnostics seed new candidate verifiers, which must beat the current verifier on the same contrasts before adoption. Across six open-ended benchmarks, UCoE averages 59.00% preference against matched direct generation with Qwen3-8B and 57.19% with Qwen3.8-27B, margins of 9.00 and 7.19 percentage points. On three verifiable benchmarks, the same method raises the mean from 36.06% for direct generation to 38.48%. Code is available at https://github.com/tianyi0216/Unified-Co-Evolution.

Abstract

Inference-Time Optimization

Explore similar work

Verifier-Harness Co-Evolution for Training-Free Self-Improvement on Open-Ended Tasks

SpecEvo: Speculative Evolution with Large Language Models for Cost-Efficient Scientific Discovery

FreeEvolve: Learning to Evolve Beyond Fixed Loops